Learn about training neural networks in TD-MPC (Temporal Difference Model Predictive Control) through a 38-minute technical video lecture that delves into the implementation details and theoretical foundations. Explore the training process by examining neural network requirements, batch structure, forward passes, and loss calculations. Understand key concepts like latent state representation stability, TD Learning principles, and their practical applications in real experiments. Master the optimization of Q networks using TD error and differentiate between offline and online data collection approaches. Access referenced research papers and implementation code from the LeRobot library while benefiting from detailed explanations across multiple topics, from basic network training to advanced concepts in temporal difference learning.
Training Neural Networks for Temporal Difference Model Predictive Control - Part 2
HuggingFace via YouTube
Overview
Syllabus
- Listing the neural networks we need to train
- What a training batch item looks like
- Forward passes and losses
- Why the latent state representation does not collapse
- Understanding TD Learning
- TD learning intuition in real experiments
- Optimizing the Q network using the TD error
- Offline vs online data collection and training loop
- Wrapping up
Taught by
Hugging Face
